Home on Eliezer de Souza da Silva

Work with me

Fri, 22 May 2026 00:00:00 +0000

Work with me: Probabilistic AI, funding routes, and mathematical rabbit holes #

I am an Assistant Professor at the University of Coimbra working on Probabilistic AI: Bayesian modelling, prior predictive analysis, GFlowNets, amortized inference, uncertainty-aware machine learning, trustworthy adaptive systems, and the occasional mathematical rabbit hole that looked harmless at first.

CV

Tue, 14 Apr 2026 00:00:00 +0000

My full curriculum vitae is available here:

Recent Accepted Papers (2026): PROPOR, AISTATS, and ICLR Workshop

Tue, 14 Apr 2026 00:00:00 +0000

I am very happy to share three recent accepted publications from 2026:

1. MATH-PT: A Math Reasoning Benchmark for European and Brazilian Portuguese #

Authors: Tiago Teixeira, Ana Carolina Erthal, Juan Belieni, Beatriz Canaverde, Miguel Faria, Diego Mesquita, Eliezer de Souza da Silva, André Martins
Venue: 17th Conference on Computational Processing of Portuguese (PROPOR 2026)

MATH-PT: A Math Reasoning Benchmark for European and Brazilian Portuguese

Sun, 01 Mar 2026 00:00:00 +0000

A benchmark of 1,729 native Portuguese math problems (European and Brazilian variants) for evaluating mathematical reasoning in modern language models.

PROPOR 2026

Orthogonal Gradient Projection for Continual LLM Unlearning

Fri, 20 Feb 2026 00:00:00 +0000

A workshop paper proposing orthogonal gradient projection for continual LLM unlearning in recursive self-improvement settings.

ICLR 2026 Workshop on AI with Recursive Self-Improvement

On the Identifiability of Tensor Ranks via Prior Predictive Matching

Thu, 15 Jan 2026 00:00:00 +0000

A principled framework for tensor rank identifiability based on prior predictive moment matching, with closed-form estimators for identifiable tensor models.

AISTATS 2026

Curso - Aprendizado Profundo / Deep Learning

Mon, 08 Sep 2025 00:00:00 +0000

Disciplina de Aprendizado Profundo / Deep Learning (MDCC UFC, Mestrado e Doutorado em Computação da Universidade Federal do Ceará) #

📅 Datas: Setembro de 2025 a Janeiro de 2026 (Semestre 2025/2)

Aulas Sobre Amostragem – Aprendizado de Máquina Probabilístico

Thu, 30 Jan 2025 00:00:00 +0000

Disciplina de Aprendizado de Máquina Probabilística (MDCC UFC, Mestrado e Doutorado em Computação da Universidade Federal do Ceará) #

📅 Datas: 30 de janeiro de 2025 e 4 de fevereiro de 2025
👨‍🏫 Professor anfitrião: César Lincoln Cavalcante Mattos (UFC)
📍 Disciplina: Aprendizado de Máquina Probabilístico
🖥️ Tema: Introdução a Métodos de Amostragem para Inferência Bayesiana

Paper Accepted at ICLR 2025 (Spotlight): When do GFlowNets Learn the Right Distribution?

Wed, 29 Jan 2025 00:00:00 +0000

I am delighted to announce that our paper, When do GFlowNets Learn the Right Distribution?, has been accepted for presentation at ICLR 2025! This work advances our theoretical understanding of Generative Flow Networks (GFlowNets) by examining the impact of balance violations on their ability to approximate target distributions and proposing a novel metric for assessing correctness.

When do GFlowNets Learn the Right Distribution?

Wed, 29 Jan 2025 00:00:00 +0000

Analysis of the limitations and stability of GFlowNets under balance violations, showing how these affect accuracy. We introduce a novel metric for assessing correctness, improving evaluation beyond existing protocols.

ICLR 2025 (Spotlight, ~top 5% 🎉)

Workshop Papers Accepted at LatinX in AI Research at NeurIPS 2024

Wed, 20 Nov 2024 00:00:00 +0000

I am thrilled to share that two of our papers have been accepted at the LatinX in AI Research Workshop at NeurIPS 2024! These works push the boundaries in causal inference and generative modeling, contributing new methodologies and insights to the field.

Paper Accepted at NeurIPS 2024: On Divergence Measures for Training GFlowNets

Sat, 28 Sep 2024 00:00:00 +0000

I’m excited to announce that our paper, “On Divergence Measures for Training GFlowNets,” authored by Tiago da Silva, Eliezer de Souza da Silva, and Diego Mesquita, has been accepted at NeurIPS 2024! 🎉

On Divergence Measures for Training GFlowNets

Fri, 27 Sep 2024 00:00:00 +0000

Novel approach to training Generative Flow Networks (GFlowNets) by minimizing divergence measures such as Renyi-$\alpha$, Tsallis-$\alpha$, and Kullback-Leibler (KL) divergences. Stochastic gradient estimators using variance reduction techniques leads to faster and stabler training.

NeurIPS 2024 (Poster)

About

Tue, 23 Jul 2024 00:00:00 +0000

Biography #

I am Eliezer de Souza da Silva, Assistant Professor at the Department of Informatics Engineering, University of Coimbra. My research centers on probabilistic AI, Bayesian machine learning, amortized inference, and trustworthy decision-making.

Research

Tue, 23 Jul 2024 00:00:00 +0000

My research program focuses on probabilistic AI: how to build learning and decision systems that remain accurate, interpretable, and reliable under uncertainty.

Thematic areas #

1) Amortized inference and GFlowNets #

I study the theory and practice of GFlowNets and related generative models, with emphasis on training objectives, stability under balance violations, and correctness assessment.

Analyzing GFlowNets: Stability, Expressiveness, and Assessment

Mon, 01 Jan 2024 00:00:00 +0000

How balance violations impact the learned distribution, motivating an weighted balance loss to improve training. For graph distributions, there are scenarios where balance is unattainable, and richer embeddings of children’s states is needed enhance expressiveness. To measure of distributional correctness in GFN we introduce a provable correct novel assessment metric.

Human-in-the-Loop Causal Discovery under Latent Confounding using Ancestral GFlowNets

Sun, 01 Jan 2023 00:00:00 +0000

We introduce a causal discovery method that estimates uncertainty and refines results with expert feedback. Using generative flow networks, we sample belief-based ancestral graphs that captures latent-confounding, and iteratively reduce uncertainty through human input, with a human-in-the-loop approach.

Prior Specification for Bayesian Matrix Factorization via Prior Predictive Matching

Sun, 01 Jan 2023 00:00:00 +0000

A method for prior specification by optimizing hyperparameters via the prior predictive distribution. This approach matches virtual statistics generated by the prior to certain target values. We apply it to Bayesian matrix factorization models, obtaining a close-formula for the rank of the latent variables, and analytically determine the matching hyperparameters, and extend it to general models through stochastic optimization.

JMLR 2023

ICML 2024 (Poster, Journal Track)

Time is of the Essence: a Joint Hierarchical RNN and Point Process Model for Time and Item Predictions

Tue, 01 Jan 2019 00:00:00 +0000

A joint model combining a Hierarchical RNN for session-based recommendations and a Point Process model for predicting return times. This approach improves both recommendation accuracy and return-time predictions over strong baselines.

WSDM 2019 (Poster)

Content-Based Social Recommendation with Poisson Matrix Factorization

Sun, 01 Jan 2017 00:00:00 +0000

A latent variable probabilistic model for recommender systems that combines social trust, item content, and user preferences into a unified Poisson matrix factorization framework. This model jointly factorizes the user–item interaction matrix and item–content matrix, accounting for social relationships and content information to enhance recommendation accuracy. ECML 2017

(old post, test) Sampling from Dirichlet Distribution using Gamma Distributed Samples

Sun, 13 Mar 2016 00:00:00 +0000

There is an algorithm to generate Dirichlet samples using a sampler for Gamma distribution for any $ \alpha > 0 $ and $ \beta > 0 $. We will generate Gamma distributed variables $ z_k \sim \text{gamma}(\alpha_k,1) $, for $ k \in {1,\cdots,d} $, and do the following variable transformation to get Dirichlet samples $ x_k = \frac{z_k}{\sum_k z_k} $. First, we should demonstrate that this transformation results in Dirichlet distributed samples.

Large-scale distributed locality-sensitive hashing for general metric data

Wed, 01 Jan 2014 00:00:00 +0000

A distributed-memory approach for Locality-Sensitive Hashing (LSH) that generalizes to metric spaces using Voronoi diagrams and enables efficient large-scale similarity search.

SISAP 2014